next up previous [pdf]

Next: Fast 3D butterfly algorithm Up: Theory Previous: Theory

Basic formulation

The right-hand side of equation 5 is a quotient of two (discrete) generalized Radon transforms (Beylkin, 1984). They can be expressed in a unified way as (to simplify the notation, we write $ p=W_{\cos}$ , $ q=W_{\sin}$ in this and next subsections)

$\displaystyle (Rg)(\tau,p,q)=\sum_{x,y}g(\sqrt{\tau^2+p(x^2-y^2)+2qxy},x,y),$ (7)

where $ g$ is $ d$ or some composite function of $ d$ .

To construct the fast algorithm, we first rewrite equation 7 in the frequency domain as

$\displaystyle (Rg)(\tau,p,q)=\sum_{f,x,y} e^{2\pi i f \sqrt{\tau^2+p(x^2-y^2)+2qxy}}\hat{g}(f,x,y),$ (8)

where $ f$ is frequency and $ \hat{g}(f,x,y)$ is the Fourier transform of $ g(t,x,y)$ in time. We next perform a linear transformation to map all discrete points in $ (f,x,y)$ and $ (\tau,p,q)$ domains to points in the unit cube $ [0,1]^3$ ; i.e., a point $ (f,x,y)\in[f_$min$ ,f_$max$ ]\times[x_$min$ ,x_$max$ ]\times [y_$min$ ,y_$max$ ]$ is mapped to $ \mathbf{k}=(k_1,k_2,k_3)$ $ \in[0,1]\times[0,1]\times[0,1]=K$ via
    $\displaystyle f=(f_$max$\displaystyle -f_$min$\displaystyle )k_1+f_$min$\displaystyle ,$  
    $\displaystyle x=(x_$max$\displaystyle -x_$min$\displaystyle )k_2+x_$min$\displaystyle ,$  
    $\displaystyle y=(y_$max$\displaystyle -y_$min$\displaystyle )k_3+y_$min$\displaystyle ;$  

a point $ (\tau,p,q)\in[\tau_$min$ ,\tau_$max$ ]\times[p_$min$ ,p_$max$ ]\times[q_$min$ ,q_$max$ ]$ is mapped to $ \mathbf{x}=(x_1,x_2,x_3)\in[0,1]\times[0,1]\times[0,1]=X$ via
    $\displaystyle \tau=(\tau_$max$\displaystyle -\tau_$min$\displaystyle )x_1+\tau_$min$\displaystyle ,$  
    $\displaystyle p=(p_$max$\displaystyle -p_$min$\displaystyle )x_2+p_$min$\displaystyle ,$  
    $\displaystyle q=(q_$max$\displaystyle -q_$min$\displaystyle )x_3+q_$min$\displaystyle .$  

If we define a phase function $ \Phi(\mathbf{x},\mathbf{k})$ as

$\displaystyle \Phi(\mathbf{x},\mathbf{k})= f \sqrt{\tau^2+p(x^2-y^2)+2qxy},$ (9)

then equation 8 can be recast as

$\displaystyle (Rg)(\mathbf{x})=\sum_{\mathbf{k}\in K} e^{ 2\pi i \Phi(\mathbf{x},\mathbf{k})}\hat{g}(\mathbf{k}), \quad \mathbf{x}\in X$ (10)

(throughout the paper, $ K$ and $ X$ are used to denote either sets of discrete points or the cubic domains containing them).


next up previous [pdf]

Next: Fast 3D butterfly algorithm Up: Theory Previous: Theory

2015-03-27